Reddit thread: agents can manufacture genuine evidence by acting on unverified claims
usefulwrong · reddit · 2026-11-17
Using ProPublica's investigation of Brownsville Energy Storage Terminals as a starting point, the author shows how Google AI Overview already repeats unverifiable claims from a shell company's website.
He then describes a subtler failure mode in multi-agent systems: one agent acts on unverified info (creates a vendor record, opens an inquiry); months later another agent treats those genuinely existing records as confirmation of the original claim. Since the new records have their own origins, source-tracing won't necessarily catch this. He asks how builders track not just where information came from, but which decisions exist because an earlier system believed it.
More from AGI Musings
- Report: US Has 9x China's Compute, So It Can Afford to Pace the AI Frontier — trevposts · 2026-11-17
- Open-source models now serve 30% of tokens, growing as fast as closed frontier AI — ivan_bezdomny · 2026-11-17
- Meme: in a few decades a 250M-param model in a comparator harness will be your judge — ctjlewis · 2026-11-17
- The 'non-technical' gap is just missing prerequisites — and Claude closes it — brandon_galang · 2026-11-17
- European game unions issue anti-GenAI manifesto as devs push back — isnit0 · 2026-11-17
- Paul Graham hails updated METR graph as 'stupendously fast' AI growth — ben_j_todd · 2026-11-17